TensorRT

Why we included this project

Torch-TensorRT compiles PyTorch models into TensorRT engines so they run faster on NVIDIA GPUs without hand-rewriting the model. Using it as a torch.compile backend gets most of the speedup with a one-line change, up to 5x lower latency than eager execution, and the export workflow serializes an optimized module you can load later from C++/libtorch in a production server. It fits vision models, diffusion pipelines, and LLM inference on data-center or Jetson hardware, and supports Linux, Windows (Dynamo only), and Jetson through source builds. The tooling handles FP8 quantization and ahead-of-time compilation, and because the PyTorch and NVIDIA teams keep it maintained, it tracks current framework versions instead of lagging behind.